May 28, 2023, Kai-Fu Lee, Chairman and CEO of Innovation Works, delivered a keynote speech at the “Artificial Intelligence Large Model Development Forum” parallel forum of the Zhongguancun Forum on May 28th, explaining new opportunities from AI 1.0 to AI 2.0. Despite the imperfect nature of large models, including instances of “hallucinations” during response generation, Lee believes these models still hold potential for enormous commercial value. He anticipates that the greatest future applications will incorporate large AI models, with “AI first” companies emerging as the favorites of the era.
Imperfect Large Models Still Bring Trillions in Commercial Value
Lee pointed out that large models will reconstruct product ecosystems and bring tremendous opportunities. The main problem with large models at present is their tendency to “bluster” while generating responses. This issue is challenging to resolve as the reasoning ability of large models comes from the same source. Reducing this “blustering” ability to near zero could risk compromising the reasoning capabilities of these models. However, there are potential solutions, such as data alignment, early warnings, corrections, or transforming large models into smaller ones in certain domains to reduce the chance of AI “blustering”.
Even an imperfect large model, according to Lee, could still have significant commercial value, potentially on a scale of trillions. “If journalists and lawyers use AI to assist in writing or drafting litigation contracts, we cannot regard it as the ultimate application due to possible errors, for which humans would still be responsible. However, in graphic design and entertainment, especially in entertainment applications, users would not mind minor errors. For instance, if a game character’s beard is a bit longer or shorter, or if a line is spoken incorrectly, it doesn’t matter because the content of the game is made up. Thus, we can tolerate these imperfect large models in many domains,” Lee said.
Lee provided examples of future possibilities, such as children in a community creating a game they want to play, using text to introduce the game and completing the game creation in a matter of seconds. In e-commerce and advertising, AI can tailor advertisements and images for each product based on the needs, cognition, educational level, and purchasing habits of each individual to increase their purchasing power. There would, of course, be regulatory issues. If the content written is false or harmful to users, legal supervision would be required. But these two examples illustrate that large models are not just question-and-answer engines; they will transform the app ecosystem across all industries.
Large Models: An Unmissable Historical Opportunity That Will Redefine Every Application
Lee believes that large models will not only reform artificial intelligence but also create enormous platform differences between companies. The ones that stand out will be those with “AI first” applications. He explained that “AI first” refers to companies that would not exist without large AI models.
Lee cited examples from the mobile internet boom, including apps like Meituan, Didi, and Douyin, which fully leveraged the features of mobile phones to achieve explosive growth in product applications. Accordingly, when choosing to start a business or invest in the mobile internet, one must select apps that cannot operate without the mobile internet. Similarly, when developing apps or starting a business in the field of AI today, one must choose apps that cannot function without large AI models. This is what is meant by “AI first”, and such apps will be the favorites of the future.
Lee is convinced that the greatest future applications will be those that cannot function without large AI models. In his view, large AI models represent an unmissable historical opportunity. They will be the largest platform ever,ten times more significant than Windows and Android, rewriting every application, restructuring human work, amplifying human intelligence tenfold or more, and replacing many repetitive tasks.
However, Lee admits that the application of this technology might have a long way to go, with practical use cases unlikely to appear in the next three to five years. “In developing large models, China faces some challenges, such as a later start compared to the US, less computing power, and less experience with large models. However, China’s market is vast, its economy is strong, and the Chinese government is better equipped to handle this work restructuring, effectively moving more Chinese people towards cooperative positions, something harder in Western countries. China has a tremendous talent advantage, with a large number of AI scientists. I believe that with the joint efforts of the government, large enterprises, start-ups, and investment companies, we will be able to quickly overcome these challenges,” Lee said.